The win rate we could have sold you

Win rate is the industry's favourite number, and it is close to meaningless. We know because we engineered a good one on purpose, looked at what it did to the returns, and threw it away. This article is that result, published in full — including the findings that are embarrassing for us, because those are the ones that tell you something.

What we actually did

We built a structure-based model of the kind sold everywhere online — liquidity sweep, shift in market structure, entry on the retracement — and tested it faithfully across five instruments over six years of data, with the most recent years locked away untouched so we could not tune against them. Fills were taken on the next bar, never the signal bar, because a backtest that fills at the price that triggered it is measuring a time machine rather than a strategy.

The 58% version

Then we tried the thing that every course teaches: split the position into three, take profit at one, one and a half, and two times risk, and move the stop to breakeven once the first one fills. It works exactly as advertised. On gold, the win rate went from 38% to 50%. On sterling, from 42% to 58%.

That is a marketable number. It is also, in our tests, the less profitable configuration. Scaling out banked the small wins and capped the large ones, while every loser still lost in full. The equity curve got smoother, the screenshots got better, and the money got worse.

Why the two numbers disagree

Win rate answers how often you are right. Expectancy answers what you earn per attempt, and it is the only one connected to your account balance. It is roughly: how often you win, times what you win, minus how often you lose, times what you lose. Nothing in that expression rewards being right more frequently if being right pays less.

This is why the levers that raise win rate are so often the ones that lower returns. All three of the popular ones did exactly that in our tests:

• Scaling out early — raises win rate, caps the winners that pay for everything else.
• Moving the stop to breakeven — feels like free risk removal. These setups routinely dip against the entry before working, so the breakeven stop mostly converted eventual winners into scratches.
• Trailing the stop — same failure, same reason. Consistently hurt expectancy across every configuration we ran.

The best-performing exit we found was the least sophisticated one: leave the stop alone and let the trade run to a fixed multiple of risk. It has a worse win rate than all three alternatives, and it made more money than all three.

The small-sample trap, which is worse

Here is the finding that should change how you read anyone's numbers, including ours. When we searched the parameter space for configurations with a win rate above 50%, we found plenty. Every single one of them had between 20 and 42 trades across six years.

Twenty-odd results is not a track record; it is a coin flipped a handful of times. When we restricted the search to configurations that actually traded — a hundred results or more — the win rate settled at 39–41%, every time. And when we loosened the entry rules further to generate more trades still, the edge inverted and went negative.

Quality was inversely related to frequency. Every configuration that looked impressive was one that had barely traded.

That relationship is the whole game. A screenshot of a 60% win rate is compatible with a genuinely good method, and it is equally compatible with someone having run a search until something landed. You cannot tell which from the number, and the number is all you are ever shown. Ask for the sample size instead. The answer, or the absence of one, is more informative than any percentage.

The rest of what we found, unflattering parts included

• The model does not generalise. Backtested across five instruments, it was positive on three and clearly negative on two. The obvious fix — tuning each instrument separately — is how you fit a strategy to the past, so we did not do it.
• The two failures were broken at the entry, not the exit. On those instruments the strike rate was so low that no exit rule rescued it. We tested a smarter one, targeting real opposing liquidity, and it was worse on all five.
• An opening-range breakout system we had tested earlier looked fine until we filled it realistically, at which point the edge disappeared entirely. We retired it rather than republish it.
• Year by year, even the profitable configurations alternated. Three up years and three down years, netting close to nothing, is not the same shape as the equity curve in the advert — and it is a shape a six-month sales window can hide completely.

Why publish this

Partly because it is true, and this whole business is a bet that saying true things about a dishonest industry is a viable strategy. But mostly because of what the result actually points at.

The mechanical version of this method is roughly break-even. Yet people do trade approaches like it profitably, and when you examine what they are doing that the code is not, it is never a secret entry rule. It is judgement applied to management: skipping the setup that technically qualifies but looks wrong, sizing down after a bad morning, letting the good one run past the target because the reason for the trade is still intact. That is the part no course sells you, because it cannot be packaged as a diagram.

Which is a slightly awkward conclusion for a company that builds trading software, and it is why our product measures behaviour rather than selling signals. The variable that decided the outcome in every test we ran was not the entry. It was what happened after.

How to read anyone's numbers, including ours

• Ask for expectancy per trade, not win rate. If only the win rate is offered, that is the answer.
• Ask how many trades. Under a hundred, treat any percentage as noise.
• Ask whether costs were included, and whether the fill was on the signal bar or the next one.
• Ask what was tested and discarded. Nobody finds a working system first try — a track record with no failures in it is a marketing document.
• Ask to see the losing years. They exist. Their absence tells you the window was chosen.